【ICLR23论文】Can CNNs Be More Robust Than Transformers?

news2024/11/15 11:07:16

文章目录

0 Abstract
1 Introduction
2 Related Works
3 Settings
- 3.1 CNN Block Instantiations
- 3.2 Computational Cost
- 3.3 Robustness Benchmarks
- 3.4 Training Recipe
- 3.5 Baseline Results
4 Component Diagnosis
- 4.1 Patchief Stem
- 4.2 Large Kernel Size
- 4.3 Reducing Activation And Normalization Layers
5 Components Combination
6 Knowledge Distillation
7 Larger Models
8 Conclusion
Acknowledgement
Reference

Article Reading Record

在这里插入图片描述

0 Abstract

Transformers are inherently (本质上) more robust than CNNs
we question that belief by closely examining the design of Transformers
simple enough to be implemented in several lines of code, namely a) patchifying（修补） input images, b) enlarging kernel size, and c) reducing activation layers and normalization layers.

1 Introduction

ViT offers a completely different roadmap—by applying the pure self-attention-based architecture to sequences of image patches, ViTs are able to attain competitive performance on a wide range of visual benchmarks compared to CNNs.

在这里插入图片描述
dubbed (被称为) vanilla（普通）

2 Related Works

Vision Transformers.
CNNs striking back ( 反击 )
ConvNeXt, shifting the study focus from standard accuracy to robustness
Out-of-distribution （分布） robustness.

ResNet Bottleneck block

counterpart （对口/对应方/同行）

在这里插入图片描述

corruption （腐败/堕落） rendition （再现） inherently （本质）

we show CNNs can in turn outperform Transformers in out-of-distribution robustness.

3 Settings

thoroughly （彻底）

3.1 CNN Block Instantiations

（实例化）

在这里插入图片描述

3.2 Computational Cost

mitigate （减轻） the computational cost loss

roughly（大约）

3.3 Robustness Benchmarks

extensively (广泛) evaluate

contains synthesized （合成） images with shape-texture（纹理） conflicting cues

image corruption（损坏）

which contains natural renditions （再现） of ImageNet object classes with different textures and local image statistics（统计）

3.4 Training Recipe

deliberately (故意) apply the standard 300-epoch DeiT training recipe

3.5 Baseline Results

we use “IN”, “S-IN”, “IN-C”, “IN-R”, and “IN-SK” as abbreviations（缩写） for “ImageNet”, “Stylized-ImageNet”, “ImageNet-C”, “ImageNet-R”, and “ImageNet-Sketch”.

4 Component Diagnosis

( 组件 )（诊断）

These designs are as follows: 1) patchifying
input images (Sec. 4.1), b) enlarging the kernel size (Sec. 4.2), and finally, 3) reducing the number
of activation layers and normalization layers (Sec. 4.3)

4.1 Patchief Stem

ViT adopts a much more aggressive down-sampling strategy by partitioning (分区) the input image into p×p non-overlapping （非重叠） patches and projects each patch with a linear layer

have investigated （调查） the importance of

when employing （使用） the 8×8 patchify stem

albeit （尽管） potentially （潜在） at the cost of clean accuracy

is boosted (提高) by at least 0.6%

play a vital（重要） role in closing the robustness gap between CNNs and Transformers.

在这里插入图片描述

4.2 Large Kernel Size

One critical（关键） property （性质） that distinguishes the self-attention operation from the classic convolution operation is its ability to operate on the entire input image or feature map, resulting in a global receptive（接收） field.

The importance of capturing long-range （远程）dependencies （依赖）has been demonstrated （证明） for CNNs even

In this section, we aim to mimic （模仿） the behavior of the self-attention block

the performance gain gradually saturates（饱和）

an unfair（不公平的） comparison to some extent.（程度）

在这里插入图片描述

4.3 Reducing Activation And Normalization Layers

(规范化层)

在这里插入图片描述

The optimal (最优) position

在这里插入图片描述

5 Components Combination

explore the impact of combining all the proposed components on the model’s performance.

along with the corresponding （相应） optimal （最优） position for placing the normalization and activation layer

An exception （异常） here is ResNet-Inverted-DW

we empirically （经验） found that using a too-large kernel size

6 Knowledge Distillation

when the model roles are switched （互换）, the student model DeiT-S remarkably outperforms the teacher model ResNet-50 on a range of robustness benchmarks

7 Larger Models

To demonstrate （演示） the effectiveness of our proposed models on larger scales
在这里插入图片描述

8 Conclusion

By incorporating （合并） these designs into ResNet,
we have developed a CNN architecture that can match or even surpass （超越） the robustness of a Vision Transformer model of comparable size.

We hope our findings prompt researchers to reevaluate（重新评估）the robustness comparison between Transformers and CNNs, and inspire further investigations （调查） into
developing more resilient （弹性） architecture designs